Comparative Analysis of ResNet for Offline Signature Verifcation Using Various Activation Functions

Signature verification is an essential function in biometric authentication, especially for safeguarding document integrity and confirming identity. This research investigates the utilization of ResNet, a deep convolutional neural network, for offline signature verification. We conduct a comparative analysis of ResNet-18 utilizing various activation functions, emphasising ReLU (Rectified Linear Unit) and ELU (Exponential Linear Unit). The selection of an activation function can profoundly affect model performance, influencing both convergence speed and total accuracy. We utilize a publicly accessible signature dataset (CEDAR) to assess the models' performance, assessing critical metrics such as accuracy. Extensive investigation reveals that selecting activation function is crucial in achieving a balance between learning capacity and generalization. Our findings indicate that ResNet-18 with the SWISH activation function surpasses both ELU and ReLU in specific contexts, especially in managing the subtleties of signature variations. This research provides significant insights into enhancing deep learning models for biometric applications, offering a framework for advancing signature verification systems.